Details of common Models in PyCBC Inference
Commonly used likelihood models are compared below. PyCBC Inference also has the capability to interface with external models. See the interactive tutorial if you’d like to learn how to add you own model for handling a new problem such as GRBs, Kilonova or whatever you can imagine.
Standard models with full waveform generation
'gaussian_noise' pycbc.inference.models.gaussian_noise.GaussianNoise
The gaussian_noise model is one of the most generic models in PyCBC Inference. It is designed for gravitational-wave data that is assumed to be Gaussian and stationary. Otherwise, there are no restricts on the types of waveform models that may be used. Waveform models supported by either get_fd_waveform or get_td_waveform can be used as waveform models. Use this model only if none other fits and is more optimized for your specific problem. Because this waveform model is generic is often slower than the other more specialized models.
Supported Marginalizations: None
'marginalized_phase' pycbc.inference.models.marginalized_gaussian_noise.MarginalizedPhaseGaussianNoise
The marginalized_phase model is a straightforward extension to the gaussian_noise model that marginalizes the signal over an overall phase. This can account for the orbital phase of a gravitational-wave signal which is usually a nuissance parameter. However, this implementation is only correct for signals that only include the dominant gravitational-wave mode (or 22 mode). This is because each mode has a different functional dependence on the orbital phase.
Supported Marginalizations: coa_phase (automatic, dominant-mode)
'marginalized_polarization' pycbc.inference.models.marginalized_gaussian_noise.MarginalizedPolarization
The marginalized_polarization model is a straightforward extension to the gaussian_noise model that marginalizes the signal over the polarization angle of the gravitational-wave signal. This is done by filtering both the plus and cross polarizations separately and then numerically integrating the resulting inner products over a grid of polarization angle values. The density of the grid is selectable. Additional marginalizations can also be optionally enabled. This model is often used in perference to the marginalized_phase model because it can support waveform models with higher order modes.
Supported Marginalizations: polarization (automatic), coa_phase (dominant-mode), distance
'marginalized_time' pycbc.inference.models.marginalized_gaussian_noise.MarginalizedTime
The marginalized_time model calculates a time series of inner products between the signal model and data so that the likelihood can be numerically marginalized over. The marginalization uses a weighted monte-carlo which samples the most likely regions of the time series more densly. The time series is also sub-sample interpolated. Higher mode waveforms are also permitted as both the plus and cross polarizations are explicitly handled separately. The time marginalization can also be done in concert with sky and various intrinsic marginalizations.
Supported Marginalizations: tc (time), distance, coa_phase (dominant mode), polarization, ra dec
'marginalized_hmpolphase' pycbc.inference.models.marginalized_gaussian_noise.MarginalizedHMPolPhase
The marginalized_hmpolphase model numerically marginalizes the likelihood over a grid of both polarization values and orbital phase values. It explicitly handles this for higher-order-mode waveforms by calculating all inner products between the data and signal on a mode-by-mode basis. The likelihoods are then assembled for each polarization and phase value and numerically integrated over. This model can only be used for waveform approximants which support the PyCBC higher mode interface as we need to be able to calculate the each mode separately.
Supported Marginalizations: polarization (automatic), coa_phase (automatic)
Heterodyne / Relative Models
'relative' pycbc.inference.models.relbin.Relative
The relative model uses a reference signal provided in its configuration to expand the likelihood in terms of differences between the reference signal and a target signal. If the reference is close to the peak in the likelihood, the reasonable portions of parameter space to explore will only have small phase deviations from the reference signal (i.e. several cycles). This allows us to represent the ration between target signal and our referencee using a piece-wise linear approximation. The target signal only needs to calculated at edges of this approximation. This model thus only supports waveforms which can efficiently generate a signal only at a given set of frequency values. Higher order modes are not recommended due their possible violation of the approximation that the ration between target and reference signals is a slowly varying smooth function. In this case use the multi_signal model and treat use a relative model for each mode. Where the model approximations hold, use this in preference to the models that need to generate a full waveform for every likelihood as these will usually be much faster.
There is also support in this model for use with LISA Sangria data analysis and LISA injection data analysis.
Supported Marginalizations: distance, coa_phase (dominant mode), polarization
'relative_time' pycbc.inference.models.relbin.RelativeTime
The relative_time model extends the relative model by using calculating the likelihood for a grid of possible merger times. A weighted monte-carlo is performed to integrate over the merger time. Sub-sample interpolation of the time series is performed. Sky location marginalization among others can be done in concert.
Supported Marginalizations: distance, coa_phase (dominant mode), polarization, ra dec
'relative_time' pycbc.inference.models.relbin.RelativeTime
The relative_time_dom model further specializes the model to completely preclude use with higher order mode signals. For this restriction, it gains the ability to marginalize over inclination
Supported Marginalizations: distance, coa_phase (dominant mode), polarization, inclination, ra dec
'brute_lisa_sky_modes_marginalize' pycbc.inference.models.relbin.Relative
The models does a brute force marginalization over the LISA sky mode degeneracies. It is built upon the relative model
Supported Marginalizations: distance, coa_phase (dominant mode)
Extrinsic Parameter Only Models
'single_template' pycbc.inference.models.single_template.SingleTemplate
The single_template model is only for extrinsic parameter estimation e.g. sky location, distance, inclination, etc. It speeds up parameter estimation by completely avoiding waveform generation while calculating likelihoods. A single reference signal is generated with fixed intrinsic (masses, spins, etc) parameters. The likelihoods can then be precalculated up to constant factors which vary with the extrinsic parameters. Only dominant-mode signals are supported as the plus and cross polarizations are assumed to be different only by a phase. With this model all supported parameters may be marginalized over or any subset.
Supported Marginalizations: tc (time), distance, coa_phase (dominant mode), polarization, ra dec
Composite Models
pycbc.inference.models.hierarchical.HierarchicalModel
The hierachical model is a container for submodels. Each submodel makes an indepedent contribution to an overall likelihood function.
'multi_signal' pycbc.inference.models.hierarchical.MultiSignalModel
This is a container for submodels where each model shares use of the same data and hence there may be cross terms between the several signals. This model requires support from the submodel to handle cross-term calculations. Some supported models include the gaussian_noise, relative and single_template.